Flatten dataset into one sample per row

Add 8,656 flattened samples as data/train.jsonl, update the Dataset Card, and preserve the original qace.json as archive/qace.json.original. No rows were removed or deduplicated.

README.md

@@ -1,15 +1,15 @@
 ---
 task_categories:
 - text-generation
-pretty_name: math
+pretty_name: Vietnamese School Math Dataset
 size_categories:
-- 100K<n<1M
+- 1K<n<10K
 ---
-# Vietnamese School Math Dataset
 
-`vi_math_school_full` is a Vietnamese-language dataset containing school-level mathematics problems and related text data.
+# Vietnamese School Math Dataset
 
-The dataset is intended for training and evaluating language models on Vietnamese mathematical understanding, problem solving, reasoning, and educational question answering.
+`vi_math_school_full` is a Vietnamese-language dataset containing 8,656
+school-level mathematics problems and related explanations.
 
 ## Main Uses
 
@@ -21,16 +21,34 @@ The dataset is intended for training and evaluating language models on Vietnames
 
 ## Language
 
-* Vietnamese
+Vietnamese
 
 ## Domain
 
-* School mathematics
-* Educational content
-* Mathematical reasoning
+School mathematics, educational content, and mathematical reasoning.
+
+## Data layout
+
+The recommended upload is one JSON object per line in `data/train.jsonl`.
+Each line is one sample with these columns:
+
+* `id`: original sample identifier. IDs are not globally unique because two
+  source groups reuse numeric IDs.
+* `question`: question text.
+* `choices`: list of answer choices; the original data contains 2–5 choices
+  depending on the sample (most samples contain 4).
+* `explanation`: solution/explanation text.
+* `answer`: answer text; one original sample has a null answer.
+
+The source data was originally stored as six nested arrays. It was flattened
+without deduplication or content rewriting so that every object is one row.
 
 ## Notes
 
-This dataset was collected and prepared primarily for experimentation with Vietnamese NLP and language models. Data quality and formatting may vary across samples, so additional cleaning or filtering may be recommended depending on the intended use.
+This dataset was collected and prepared primarily for experimentation with
+Vietnamese NLP and language models. Data quality and formatting may vary across
+samples, so additional cleaning or filtering may be recommended depending on
+the intended use.
 
-Users should independently verify generated mathematical answers when using models trained on this dataset in educational or high-stakes applications.
+Users should independently verify generated mathematical answers when using
+models trained on this dataset in educational or high-stakes applications.